On 24/08/2014 13:29, Adam Wulkiewicz wrote:
Ion Gaztañaga wrote:
An interesting addition to Boost.Container? ;-)
Some time ago I tested a container which should also improve the cacheing, depending on the access pattern. The key point is to separate the hot and cold data so its purpose is slightly different than poly_collection but the idea is similar, to group the same objects together. The idea is to transform a vector of tuples into a tuple of vectors and store tuple's components in separated containers, e.g.:
std::vector<std::tuple<T1, T2>> -> std::tuple<std::vector<T1>, std::vector<T2>>
This is AoS -> SoA conversion. We also do that kind of thing with NT2 containers (this is necessary for vectorization).
The usage should be as close to std::vector as possible, so e.g.:
|tuple_vector<std::tuple<T1, T2, T3, ...>> v; v.resize(count);
for (int i = 0; i < count; ++i) { T1 t = std::get<0>(v[i]); std::get<0>(v[i]) = ...; }|
A thing worth noticing is that the tuple_vector's reference type is a tuple of references:
std::tuple<T1&, T2&>
Could also be any proxy type, that's more flexible. The important aspect is that tuple_vector's reference cannot be value_type&.